Markov Random Field for Image Concept Detection

Abstract

Recent years have witnessed phenomenal growth in the number of internet multimedia documents, such as un-annotated scene images. Image concept detection is an important step for image query. In this paper, we tackle the problem of image concept detection, namely automatically detecting concepts in an un-annotated image. In fact, a scene image always contains multiple concepts. In order to detect all concepts in an image, traditional approaches only leverage single-concept detectors to detect them one by one, which may be ineffective in some cases. For example, the visual appearance described by a scene image consisting of a few concepts (e.g. lion, grassland, Africa) may be difficult to distinguish solely by single-concept detectors. We propose a novel Markov random field for Image Concept Detection (MICD) model, which considers modeling the link of single-concept detectors and multi-concept detectors to enhance the precision of image concept detection. Single-concept detector can recognize single concept in an image, while multi-concept detector can effectively distinguish the multi-concept scene. Their fusion can lead to the improvement of detection precision. Additionally, the interdependencies between concepts are utilized to improve detection precision. In order to investigate the feasibility and effectiveness of our model, we conduct experiments on Corel and NUS-Wide collection. The experimental results demonstrate the effectiveness of our proposed approach.